基于YOLOv7的通用目标检测模型  

A general object detection algorithm based on YOLOv7 model

作　　者：钟玲[1] 陆国芳 ZHONG Ling;LU Guofang(School of Software,Shenyang University of Technology,Shenyang 110870,China)

机构地区：[1]沈阳工业大学软件学院,沈阳110870

出　　处：《智能计算机与应用》2023年第12期23-31,共9页Intelligent Computer and Applications

基　　金：国家自然科学基金(61540069);辽宁省教育厅科研基金(LJGD2020017)。

摘　　要：针对通用目标检测领域在自动提取特征的过程中会提取错误的目标检测区域信息,本文以YOLOv7模型作为基线模型进行改进,有效地提高检测精度。首先,在YOLOv7模型的主干网络中引入改进的注意力机制,在上采样模块中采用双三次插值,以增强浅层和深层的特征融合效果,减少区域信息丢失;其次,通过设计动态IOU阈值实现动态非极大值抑制,解决固定阈值导致检测边界框冗余的问题,提升准确性;最后,采用剪枝算法对网络模型进行轻量化处理,并使用深度可分离卷积替换原始卷积。实验结果显示,本文模型在数据集上的准确率、F1值和召回率均高于其他模型,说明本文建立的基于YOLOv7模型改进的通用目标检测算法的有效性。In response to the generic target detection domain that extracts wrong target detection region information in the process of automatic feature extraction,this paper uses YOLOv7 model as the baseline model.An improved attention module is introduced in the backbone network of YOLOv7,and the upsampling algorithm is changed by using dual triple interpolation in the upsampling module;meanwhile,an optimized non-maximum suppression(NMS)method is implemented in detection,and dynamic IOU thresholds are designed to achieve dynamic NMS,which solves the problem of redundancy of detection bounding boxes due to fixed thresholds and reduces the false alarm rate;finally,a pruning algorithm is used to network Finally,the pruning algorithm is used to lighten the network model and replace the original convolutional model with depth-separable convolution.The experimental results show that the Acc,F1 values and recall rates of the model in this paper are higher than those of other models on the data set,which can illustrate the effectiveness of the improved YOLOv7-based general-purpose target detection model established in this paper.

关 键 词：YOLOv7模型 通用目标检测 注意力机制 双三次插值 剪枝算法 深度可分离卷积 

分 类 号：TP392[自动化与计算机技术—计算机应用技术]